Propagation measurement based study on relay networks

Transcription

1 TECHNISCHE UNIVERSITÄT ILMENAU Fakultät für Elektrotechnik und Informationstechnik der Technischen Universität Ilmenau Propagation measurement based study on relay networks M. Sc. Aihua Hong Dissertation zur Erlangung des akademischen Grades Doktor-Ingenieur (Dr.-Ing.) Anfertigung im: Fachgebiet Elektronische Messtechnik Institut für Informationstechnik Fakultät für Elektrotechnik und Informationstechnik Gutachter: Prof. Dr.-Ing. habil. Reiner S. Thomä (TU-Ilmenau) Prof. Dr.-Ing. Wolfgang Koch (Universität Erlangen-Nürnberg) Dr.-Ing. Egon Schulz (Huawei Technologies Duesseldorf GmbH) Vorgelegt am: Verteidigt am: urn:nbn:de:gbv:ilm

2

3 ABSTRACT Considering technological bases of next generation wireless systems, it is expected that systems can provide a variety of coverage requirements to support ubiquitous communications. To satisfy the requirements, an innovative idea, integrating network elements with a relaying capability into cellular networks, is one of the most promising solutions. The main topic of this dissertation is a propagation measurement based study on relay networks. The study includes three parts: channel modeling, performance evaluation, and verification. First of all, an empirical channel model for relay networks is proposed based on statistical analyses of measurement data. Then, advanced techniques for the throughput improvement and interference cancellation are proposed for Multiple Access Relay Networks (MARN) which are used as an example of relay networks. The performance of the considered MARN is evaluated for Rayleigh channels, and then verified for realistic channels, obtained from measurement data and from the experimental relay channel model as well. For relay channel modeling, the long-term correlation properties between links are of crucial importance due to the meshed-network topology. Although, there is a wide variety of research results for Multiple-Input Multiple-Output (MIMO) channel modeling available, the characterization of correlation properties has been significantly simplified or even completely ignored which motivates this research to be performed. In this dissertation, the experimental results of the correlation properties of Large Scale Parameters (LSP) are presented through the analysis on the real-field measurement data for both the urban and indoor scenarios. furthermore, the correlation properties have been fully introduced into the WINNER channel Model (WIM) for realistic relay channel simulations. As a further contribution of this dissertation, various advanced techniques are proposed for MARN in the presence of Unknown Interference (UKIF). Multiple Access Coding (MAC) is introduced as a multiple access technique. The use of MAC provides the signal separability at the receiver and improves throughput. Thereafter, high system resource efficiency can be achieved through relay protocol design. At the receiver, Minimum Mean Square Error (MMSE)-based spatial filtering is used to suppress UKIF while preserving multiple Mobile Station (MS)s MAC-encoded signal structure. Furthermore, an error detection aided signal selection technique is proposed for diversity increasing. The theoretical system performance with aforementioned techniques is simulated for Rayleigh channels. Thereafter, realistic channels are exploited for the performance verification. The gap between the theoretical performance and the realistic performance indicates that the assumptions made to the simplified Rayleigh-channels do not fully hold in reality. For the future relay system design, this work provides valuable information about the performance evaluation of relay networks in consideration of the correlation properties between links.

4

5 KURZFASSUNG Von der nächsten Generation von Mobilfunksystemen erwartet man eine umfassende Versorgung mit breitbandigen Multimediadiensten. Um die dafür erforderliche flächendeckende Versorgung mit hohen Datenraten zu gewährleisten, können Relay-Netzwerke einen wesentlichen Beitrag liefern. Hierbei werden Netzwerkstationen mit Relay-Funktionalität in zellulare Netzwerke integriert. Diese Dissertation befasst sich mit der Untersuchung Relay-basierter Netzwerke unter Verwendung von Ausbreitungsmessungen. Die Arbeit deckt Fragen zur Kanalmodellierung, Systemevaluierung bis hin zur Systemverifikation ab. - Zunächst wird ein auf Funkkanalmessungen beruhendes experimentelles Kanalmodell für Relay-Netzwerke vorgestellt. Im Weiteren werden technische Verfahren für Mehrfachzugriffs-Relay-Netzwerke MARN diskutiert. Die erreichbare Systemleistung wurde unter Verwendung von Rayleigh-Kanälen innerhalb einer Systemsimulation bestimmt und im Anschluss mit realen Kanälen, die sowohl direkt aus Funkkanalmessungen als auch indirekt aus dem bereits erwähnten Kanalmodell abgeleitet wurden, verifiziert. Bisherige Arbeiten zur Modellierung breitbandiger Multiple-Input Multiple-Output (MIMO) Kanäle berücksichtigen nicht oder nur sehr stark vereinfacht die Langzeitkorrelationseigenschaften zwischen den Links und werden damit der vermaschten und räumlich weit verteilten Topologie von Relay-Netzwerken gerecht. In der vorliegenden Dissertation erfolgte daher eine experimentelle Untersuchung zu den Korrelationseigenschaften von Large-Scale-Parametern LSP, die unter Verwendung von Funkkanalmessdaten aus urbanen Umgebungen und aus Innenräumen abgeleitet wurden. Die Ergebnisse hierzu fanden Eingang in das vom WINNER-Projekt entwickelte Kanalmodell. Sie erlauben damit eine realistischere Simulation von Relay-unterstützten Netzen. Einen weiteren Schwerpunkt dieser Arbeit stellen technische Verfahren dar, die eine Erhöhung der Systemleistung in MARN mit unbekannter Interferenz UKIF versprechen. Im Einzelnen handelt es sich um die Mehrfachzugriffs-Kodierung MAC - die eine verbesserte Signaltrennung auf der Empfängerseite und eine Erhöhung des Datendurchsatzes erlaubt, den Entwurf eines Relay-Protokolls zur Erhöhung der Systemeffizienz, einen Minimum Mean Square Error (MMSE) Algorithmus zur Unterdrückung unbekannter Interferenzen bei Erhaltung der MAC-Signalstruktur mehrerer Mobilstationen MS, und ein fehlererkennungsbasiertes Signalauswahlverfahren zur Diversitätserhöhung. Die vorgenannten Verfahren werden in einer Systemsimulation zunächst mit Rayleigh-Kanälen evaluiert und demonstrieren die erzielbare theoretische Leistungssteigerung. Die Berücksichtigung realer Funkkanäle innerhalb der Systemsimulation zeigt allerdings, dass die theoretische Systemleistung so in der Realität nicht erreichbar ist. Die Ursache hierfür ist in den idealisierten Annahmen theoretischer Kanäle zu suchen. Für die Entwicklung künftiger Relay-Netzwerke bieten die in dieser Arbeit aufbereiteten Erkenntnisse hinsichtlich der Langzeitkorrelationseigenschaften zwischen den Links einen wertvollen Beitrag für die Abschätzung ihrer Systemleistung auf der Basis eines verbesserten Kanalmodells.

6

7 ACKNOWLEDGMENTS The valuable discussion with various people has made this dissertation to be completed. First, I want to thank Prof. Dr.-Ing. Reiner S. Thomä, the leader of my lab, for his continuous support and encouragement, both as a supervisor and as a colleague. Further, I want to thank all the colleagues in the lab, Elektronische Messtechnik at Technische Universität Ilmenau, among whom I would like to mention especially Gerd Sommerkorn, Milan Narandžić, Dr. Wim Kotterman, and Marcus Grossmann who helped me both in practical matters, such as preparing, organizing, and performing measurement campaign, and in scientific matters, such as discussion on theoretical stuff. I am deeply indebted to Prof. Tad Matsumoto from University of Oulu, Finland, not only for fruitful technical discussion, but also for all kinds of support and encouragement in the work. During his stay at Technische Universität Ilmenau and my stay at center for wireless communication of University of Oulu, Finland, many new ideas have been stimulated and the interesting results have been published. This work has been partially sponsored by the radio system technology department of Nokia- Siemens Networks GmbH & Co. KG in Munich, Germany. My special appreciation is given to Dipl. Wolfgang Zirwas from this department for the valuable discussion and his reviewing of my work. Furthermore, I would like to thank Prof. Dr.-Ing. Wolfgang Koch from Universität Erlangen- Nürnberg and Dr. Egon Schulz from Huawei Technologies Duesseldorf GmbH for reviewing my thesis. I also would like to thank Markus Landmann, Giovanni Del Galdo and Jörg Lotze for many useful tips in terms of applying LATEX to write this thesis. My deepest gratitude goes indeed to my husband, Jian Shen, who always encouraged me to finish my thesis, and to my son, Yuchen Leon Shen, who opens a new chapter in my life. Finally, I would like to thank all my friends who are a very important part of my life.

8

9 CONTENTS ix CONTENTS Title Abstract Kurzfassung Acknowledgments Contents i iii v vii ix List of Figures xiii List of Tables xvii 1. Introduction Problem statement Contributions and overview Main contributions Overview Channel modeling methodology of relay networks Background and own contributions Background and state-of-the-art Own contributions Channel modeling process MIMO channel modeling approaches Channel modeling process for relay networks System requirements Comparison among the state-of-the-art MIMO channel models Modeling goals The definition of Large Scale Parameters (LSP) and their correlation properties Usage of the correlation models of LSP Experimental channel evaluation and modeling Background and own contributions Background and state-of-the-art Own contributions

10 x CONTENTS 3.2 Propagation scenarios Urban scenarios U1 scenario: urban micro-cell U2 scenario: urban macro-cell Indoor scenarios I1 scenario: indoor corridor Relay measurement campaigns Broadband radio channel sounding: technique and equipment Micro-cell to macro-cell urban scenario: U1 and U Indoor corridor scenario: I Extraction procedure of LSP and their correlation properties Extraction procedure of the LSP values Extraction algorithms of the LSP values The impact of the setting parameters on the LSP values Different noise cut levels Different lengths of a Space Time (ST) averaging window and different overlapping ratios Extraction methodology of the correlation models of LSP Experimental results Micro-cell to macro-cell urban scenarios (U1 and U2) Statistical distribution of LSP The intra-site correlation of the transformed LSP The inter-sector correlation of the transformed LSP The inter-site correlation of the transformed LSP Indoor corridor scenario I Statistical distribution of LSP The intra-site correlation of the transformed LSP The inter-sector correlation of the transformed LSP The inter-site correlation of the transformed LSP Main findings System structure of cooperative Multiple Access Relay Networks (MARN) Background and own contributions Background and state-of-the-art Own contributions Simulation assumptions System model of MARN Forwarding strategies The Amplify-and-Forward (AF) strategy The Decode-and-Forward (DF) strategy Proposed protocols Relaying protocol design

11 CONTENTS xi Direct transmission without relaying Always relaying protocol Adaptive relaying protocol with limited feedback Multiple Access Coding (MAC) Symbol-wise super-positioning Re-transmission at the second time-slot MMSE detection JU MMSE detection JU MMSE A-criterion JU MMSE H-criterion ML detection Error detection aided signal selection Numerical simulation results of an example scenario MAC for the direct transmission JU MMSE for the interference cancellation The functionality of the RS: the AF relay and the DF relay Error detection aided signal selection Relay protocol with limited feedback Performance verification of cooperative MARN Background and own contributions Background and state-of-the-art Own contributions Simulation channels Relay measurement campaign Application of raw measurement data for the performance evaluation Application of the relay channel model for the performance evaluation Simulation scenarios and parameters Measurement data based performance verification Impacts of the small scale propagation phenomena Small Scale Fading (SSF) Influence of the antenna element spacing at the Base Station (BS) Impacts of the large scale propagation phenomena Large Scale Fading (LSF) Path Loss (PL) plus LSF Power allocation schemes Relay channel model based performance verification Positioning of the Relay Station (RS) Conclusions Appendix A. Comparison among the state-of-the-art Multiple-Input Multiple-Output (MIMO) channel models

12 xii CONTENTS Appendix B. Examples of MAC Appendix C. Solution of the Minimum Mean Square Error (MMSE) criterion Appendix D. Glossary of operators, symbols, and acronyms D.1 Mathematical operators D.2 Acronym D.3 List of frequently used symbols Bibliography Theses Erklärung

13 LIST OF FIGURES xiii LIST OF FIGURES 1.1 The relationship of the Signal-to-Noise Ratio (SNR) values between a direct link and a dual-hop relay link, assuming that they provide the same end-to-end Shannon capacity a) A single-hop cell; b) A two-hop cell for coverage extension; c) A two-hop cell for capacity improvement; d) A two-hop cell for transmit power minimization A two-hop cell for service provision in a shadowed area The meshed structure of cluster-wise two-hop relay networks Channel modeling process The top-down modeling philosophy of the WINNER channel Model (WIM) A link level channel modeling over a Local Stationary Area (LSA) A system level channel modeling Examples of the inter-site correlation An example Power Delay Profile (PDP) and the corresponding channel parameters PDP with a direct Line of Sight (LOS) path and the remaining Multipath Component (MPC)s The intra-site auto-correlation The inter-sector correlation A hand over scenario Received power difference (in [db]) from two BSs with different PL exponents Handover probability Regeneration process and algorithms based on the correlation models of LSP Geometric distribution of Mobile Station (MS)s and BSs Distribution of the LSP values over the whole area with the inter-site correlation ρ BS and the de-correlation distance d decorr as parameters MIMO channel sounder switching scheme Antenna arrays used for a high resolution parameter estimation Ilmenau downtown seen from the 16 [m] BS antenna array Measurement routes in a 2-dimensional Ilmenau city map Photos of the Helmholtzbau measurement environment Measurement routes in the floor plan of the Helmholtzbau Getting a Channel Transfer Function (CTF) from the raw data The structure of measurement data (left) and the dimension of a CTF (right) Snapshot groups over the whole measurement data

14 xiv LIST OF FIGURES 3.10 The extraction procedure of the LSP values The first method to derive a PL model and the LSF values The second method to derive the LSF values Impact of a noise cut level on the Cumulative Distribution Function (cdf) curves of PL and Delay Spread (DS) in the LOS and None Line of Sight (NLOS) propagations Impact of a ST averaging window on the cdf curves of PL and DS in the LOS and NLOS propagations Impact of the overlapping ratio on the cdf curves of PL and DS in the LOS and NLOS propagations Unified modeling and reproduction procedures The cdf curves of LSP from the measurement data in the LOS and NLOS propagations with different BS heights An example to show the transform process of K-factor: the original cdf distribution the transformed cdf distribution the modeled cdf distribution An example to show the intra-site auto-correlation of the transformed LSF: the original intra-site auto-correlation model with an exponential decaying function with the de-correlation distance being 3.9 [m] The intra-site auto-correlation of the transformed LSP in the U1 and U2 scenarios in the LOS and NLOS propagations The dependence of the inter-site correlation of the transformed LSP on d diff,θ in the urban U1 and U2 scenarios The dependence of the inter-site correlation of the transformed LSP on d diff,θ,d BS in the I1 scenario Relay network with various structures The system structure with N blocked desired MSs having a single antenna, L Unknown Interference (UKIF) MSs having a single antenna, and one BS having n Rx antenna elements Relay transmission in two phases Flow-chart of the relay protocol with Automatic Repeat-reQuest (ARQ) Achievable capacity region of a 2-MS access channel Transmission time schemes of a multiple access channel with/without a half-duplex relaying transmission Receive structure at the BS with n Rx antenna elements The Bit Error Rate (BER) and throughput performances of MAC with a) uncoded with n Rx = 1 and L = 0, b) MAC-encoded with n Rx = 1 and L = 0, and c) MAC-encoded with n Rx = 2 and L = The BER performances of MAC with H-criterion and A-criterion in the case of n Rx = 2 and L = The BER performances of MS1 and MS2 as well as MS3 with the AF relay and the DF relay The BER performances of MS1 and MS2 as well as MS3 with the AF relay and the DF relay

15 LIST OF FIGURES xv 4.12 The BER performance of MS1 and MS2 as well as MS3 with an error detection aided signal selection technique in the case of the AF relay (left) and the DF relay (right) The BER performances of MS1 and MS2 (left) as well as MS3 (right) with the re-transmission technique Relay measurement routes in a 2-dimensional Ilmenau city map The cdf curves of SSF of the stationary BS-RS link with four polarization pairs MS s normalized received powers from BS5/BS6/RS2 along route 1 and route The cdf curves of SSF of the stationary BS-RS link based on the relay channel simulations if the RS is located in the middle of the BS and MSs The cdf curves of the BER performances of MS1 and MS2 (left) as well as MS3 (right) under the assumption that SNR sd = 0 [db] while SNR sr = SNR rd = 10 [db] The cdf curves of the fading amplitude of the end-to-end channel of a tandem relay link The cdf curves of the BER performances of MS1 and MS2 as well as MS3 based on the BS5 s measurement data (left) and based on the BS6 s measurement data (right) with different BS antenna element spacing under the assumption that SNR sd = 0 [db] while SNR sr = SNR rd = 10 [db] Dependence of the BER performances of MS1 and MS2 as well as MS3 on the LSF value along route The cdf curves of SNR sr and SNR sd as well as SNR rd along route 1 (left) and route 2 (right) based on the BS5 s and BS6 s measurement data The cdf curves of the BER performances of MS1 and MS2 as well as MS3 along route 1(left) and route 2(right) The BER performances of MS1 and MS2 (left) as well as MS3 (right) along route 2 based on the BS6 measurement data The cdf curves of the BER performances of MS1 and MS2 (left) as well as MS3 (right) with different RS positioning schemes

16

17 LIST OF TABLES xvii LIST OF TABLES 2.1 Categories of MIMO channel modeling approaches Hand-over parameter setup Propagation scenarios for relay network applications The key technical parameters of the MIMO channel sounder equipments TU-Ilmenau relay measurement campaigns: U1, U2, and I The median value and the variance of the transformed LSP in the U1 and U2 scenarios in the LOS and NLOS propagations The de-correlation distance of the transformed LSP in the U1 and U2 scenarios in the LOS and NLOS propagations The intra-site cross-correlation of the transformed LSP in the U1 and U2 scenarios in the LOS and NLOS propagations The inter-sector correlation of the transformed LSP in the U1 and U2 scenarios The inter-site correlation of the transformed LSP in the urban U1 and U2 scenarios The median value and the variance of the transformed LSP in the I1 scenario in the LOS and NLOS propagations De-correlation distance of the transformed LSP in the I1 scenario in the LOS and NLOS propagations The intra-site cross-correlation of the transformed LSP in the I1 scenario in the LOS and NLOS propagations The inter-sector correlation of the transformed LSP in the I1 scenario The inter-site correlation of the transformed LSP in the I1 scenario Simulation parameters Simulation parameters for performance verification B.1 Examples of MAC for 4-MS case after computer searching B.2 Examples of MAC for 2-MS case and 3-MS case after computer searching

18

19 1. INTRODUCTION 1 1. INTRODUCTION To satisfy rapid demands in providing a wide variety of wireless communications, it is of crucial important to support ubiquity in various service requirements such as data rate and coverage while utilizing the frequency spectrum as efficiently as possible. These requirements impose unprecedented challenges to wireless communication engineers who are mainly focusing on single-hop structured networks, both in cellular networks, such as Global System for Mobile Communications (GSM) and Universal Mobile Telecommunications System (UMTS), and in local networks like Wireless Local Area Networks (WLAN). In single-hop networks, a MS communicates directly with a BS or an Access Point (AP) 1 [13]. This single-hop structure limits the network performance. First of all, with higher operation frequency, the coverage provided by a single BS is limited to the order of several hundred meters because of PL [14] [15] [16]. Furthermore, it happens very often that a transmit signal will be heavily attenuated due to shadowing because of relatively large obstacles like tall buildings, big trees or cars in outdoor scenarios, or like walls and floors in indoor scenarios. Due to these facts, ubiquitous coverage can be achieved by increasing the geographical density of BSs and/or by increasing the transmit power. However, the former solution is too costly (lease BS positions, cabling, and maintain BSs) to be accepted. The latter solution imposes increased system interference which in turn impairs the performance. An innovative idea, integrating network elements with a relaying capability into conventional singlehop structured networks [17], is one of the most promising solutions. In relay-enabled networks, RSs are introduced into the communication between a BS and a MS having a weak link to the BS or even being blocked from the BS. The RSs receive first a signal from the BS and then forward the signal to the MS and vice versa [18]. Either BS or MS can play the roles of RS, depending on applications. Furthermore, the signal forwarding performed at a RS works in a half-duplex mode due to the self-interference. This indicates, a RS will not perform simultaneous receiving and transmitting at the same frequency/time/code/antenna due to the large difference between the incoming and outgoing signal power levels. Due to the half-duplex operation at a RS, the end-to-end capacity of a relay link is divided by the number of hops n hop. It implies that the link capacity at each hop in relay networks should be at least n hop times as large as the single-hop capacity so that the same end-to-end capacity can be achieved. Figure 1.1 shows the relationship of the SNR values between a direct link and a dual-hop relay link, assuming that they provide the same end-to-end Shannon capacity. The gap between the dashed curve and the solid curve defines the minimal SNR gain that a dual-hop relay link should provide over a direct link. A relay link outperforms a direct link in the area under the solid curve. Otherwise, a direct link is preferable. Although the end-to-end capacity of a relay link is limited due to the half-duplex operation of a 1 In this dissertation, there is no clear difference between a BS and an AP. The terminology BS is used throughout this dissertation standing for both BS and AP for the simplicity.

20 2 1. INTRODUCTION SNR: direct link [db] SNR: dual hop relay link [db] Fig. 1.1: The relationship of the SNR values between a direct link and a dual-hop relay link, assuming that they provide the same end-to-end Shannon capacity RS, the performance of networks, comprising RSs with a relaying capability, can be improved in the following senses [19] [20]. 1. Extend coverage In future mobile communication systems, the coverage of a single BS becomes smaller due to the fact that a higher carrier frequency is expected to be allocated. As a trade off between providing wide spectrum of services in wide area and avoiding costly expenditure due to the deployment of huge amounts of BSs, RSs will be the most promising solution. At the coverage margin of a BS, several RSs can be deployed to extend the coverage of a single BS as shown in subplot b) in Fig Compared with a single-hop cell, the coverage of a two-hop cell becomes larger. The high link quality between the BS and the RS can be supported by using directional antenna with high antenna gain at the both sides. This type of RS application is especially efficient to the area where coverage is much more crucial than Quality of Services (QoS), for example, suburban and rural area with a low density of population. 2. Capacity improvement When the coverage is not a crucial issue, the relay concept can be used to improve the system capacity or spectrum efficiency. In a cell, RSs may be deployed in a relatively nearby area around the BS as shown in subplot c) in Fig The coverage of a two-hop cell is as the same as that of a single-hop cell. However, the power distribution in the outer area between them is different. A two-hop cell has high power over the whole cell area. As a consequence, the cell capacity can be improved. Furthermore, the fairness between the MSs having different distance to the BS can be improved. In a single-hop cell, the MSs near to the BS can normally achieve better QoS than the MSs far from the BS due to the fact that the received power is normally decreased with an increasing distance between the MS and the BS. However, in a two-hop cell, MSs can still have a good QoS even in the case that they are not in the near field of a BS. This kind of RS application is especially attractive for the hot spot areas, such as campus and airport as well as center railway station.

21 1. INTRODUCTION 3 BS RS a) b) c) d) Fig. 1.2: a) A single-hop cell; b) A two-hop cell for coverage extension; c) A two-hop cell for capacity improvement; d) A two-hop cell for transmit power minimization 3. Transmit power reduction When a system aims neither at coverage extension, nor at capacity increasing, a two-hop cell can be used to minimize the transmit power both at BSs and at MSs as shown in subplot d) in Fig The obvious consequences are, firstly, the life longevity of a battery at MSs will be larger; secondly, the system interference will be reduced. 4. Shadowing Mitigation In several applications, the receive signal from a BS is significantly attenuated due to shadowing. The shadowed area could be covered by using RSs. In outdoor urban scenarios, for example, the coverage of a BS can be extended to the shadowed area if a RS is deployed around the street corner as shown in Fig Fig. 1.3: A two-hop cell for service provision in a shadowed area Furthermore, various advanced techniques, such as cooperative relaying [21] [22] [23] [24] [25] [26] and distributed MIMO [27] as well as bi-directional relaying [28], have been proposed in relayenabled networks. Significant performance improvement has been observed. These advanced techniques benefit from the meshed network topology of relay-enabled networks which is different from

22 4 1. INTRODUCTION the star-topology of single-hop networks. Figure 1.4 shows an example of the meshed structure of cluster-wise two-hop relay networks. This system includes three clusters of stations, the BS cluster, the RS cluster, and the MS cluster. A meshed connection among clusters forms a distributed MIMO system even in the case that each station has a single antenna. In this distributed MIMO system, the number of transmit/receive antennas (n Tx /n Rx ) is the sum of antennas of all stations within one cluster. The meshed multiple BS-RS-MS system has the potential to take maximum advantage from MIMO techniques [29] [30]. BS1 MS1 RS2 BS2 MS3 RS1 BS3 MS2 BS cluster RS cluster MS cluster Fig. 1.4: The meshed structure of cluster-wise two-hop relay networks 1.1 Problem statement In literature [21] [22] [23] [24] [25] [26] [27] [28], it has been stated that relay-enabled networks are one of the most promising techniques for the future wireless communication system design. The cooperative relay concept, based on the meshed network structure of relay-enabled networks, can further improve the system performance. The performance of cooperative relay networks in a meshed link scenario will heavily depend on the correlation properties between meshed links. That does not only include the pure availability of the links. It will also include statistical space-time parameters and the interdependence of these parameters between different links of meshed relay networks as well as their temporal and spatial properties. However, majority of the current research activities about relay topic are based on analytical channel models [31] [32] [33] or existing channel models [34] [35] [36] [1]. Both the analytical channel models and majority of the existing channel models focus on describing the short-term/long-term wave propagation between two stations. An interesting observation from the literature is that the correlation properties between meshed links are either out of discussions or significantly simplified. Therefore, for the purpose of the accurate performance evaluation of relay networks, the current knowledge about the realistic broadband relay channels is rare and insuffi-

23 1.2 CONTRIBUTIONS AND OVERVIEW 5 cient. Thus, it is necessary to study the relay channels and then provide the channel models which support more realistic relay network simulations than conventional very simple ones. Due to the limited knowledge about relay channels, majority of the research works in the performance evaluation of relay networks have been performed under various simplified assumptions on the channel properties. These assumptions include for example, Rayleigh-distributed fading, an exponential PL model without LSF, and the independent channel fading between different distributed links. As a consequence, the simulation results are attractive but unrealistic. Therefore, in order to provide a realistic insight into the performance of future relay networks and to give valuable information to the future relay network deployment, the theoretical performance should be verified based on the realistic relay channels, either from real-field measurement campaigns directly or from the channel model supporting a relay structure. These topics can be detailed in the following crucial questions: 1. What are the main challenges of the channel modeling for relay networks, comparing with the state-of-the-art channel models? 2. What are the key channel metrics for the relay channel modeling? 3. How can we obtain the experimental results for the relay channel modeling, based on measurement data? 4. What kind of techniques can be proposed to improve the system performance of MARN which are used as an example of relay networks? 5. What is the realistic performance of MARN, both based on measurement data and based on the relay channel model? 1.2 Contributions and overview Main contributions The main contributions of this dissertation can be summarized as follows. First of all, the correlation properties of LSP are highlighted as the most important channel metrics for the channel modeling for relay networks. Based on relay measurement campaigns, this dissertation provides the experimental results of these channel metrics and addresses the method of how to acquire the results. Furthermore, based on the WIM, an experimental relay channel model is proposed. Thereafter, various advanced techniques for the throughput improvement and interference cancellation are proposed to an example of relay structure: MARN. The achievable system performance with these techniques is evaluation over the simplified Rayleigh channel model and then verified through realistic channel realizations derived both from raw measurement data and from the aforementioned empirical relay channel model.

24 6 1. INTRODUCTION Overview This subsection describes roughly the contributions of this dissertation chapter by chapter and how the dissertation is organized. A detailed description to the contributions of each chapter could be found at the beginning of each chapter. Both Chapter 2 and Chapter 3 deal with the channel modeling for relay networks. Chapter 2 discusses the channel modeling methodology while Chapter 3 presents the experimental results for both urban and indoor scenarios through the analysis on the measurement data gathered from real-field measurement campaigns. The author s contributions related to some of the topics of these two chapters have been published in [1] [2] [3] [4] [5] [6] [7] [8] [9]. Chapter 2 describes firstly the general channel modeling process and the MIMO channel modeling approaches. Thereafter, Chapter 2 specifies the channel modeling process of relay networks. In this process, at first, the system requirements of relay networks are analyzed. Then, the state-of-the-art MIMO channel models are compared, by which the WIM is selected as the initial channel model of relay networks. Thereafter, channel modeling goals for a full relay networks support are discussed based on the WIM framework. The main goal is modeling the correlation properties of LSP. At the end of this chapter, the method of how to reproduce relay channel realizations is presented, by using the correlation models of LSP. Chapter 3 presents the experimental results of the correlation properties of LSP, by statistically analyzing the measurement data gathered from two scenarios: urban micro/macro scenarios and indoor scenario. The detailed information about the measurement campaigns and measurement scenarios is provided at the beginning of this chapter. Thereafter, the extraction procedure of the LSP values and their correlation properties is described. Whereby, the impact of the setting parameters on LSP is investigated. At the end of this chapter, experimental results are presented. The results include the statistical distribution of LSP, the intra-site correlation, the inter-sector correlation, and the inter-site correlation of LSP. Chapter 4 studies the theoretical performance of MARN which are used as an example of relay networks. First of all, this chapter shows the general system model of MARN in the presence of UKIF. Two different forwarding strategies: an AF relaying and a DF relaying, have been under consideration. Based on this system model, advanced techniques, such as MAC [37], relay protocol design [25] [22], a symbol-wise super-positioning, and the Joint User (JU) MMSE algorithm [38] as well as an error detection aided signal selection technique [22] [39], are proposed for the further system performance improvement. Significant performance enhancement has been shown at the end of this chapter based on the numerical simulations, using the simplified Rayleigh channel model. Own contributions related to some of the topics of this chapter have been published in [10] [11] [12]. The numerical results of MARN with the proposed advanced techniques have been highlighted in Chapter 4 based on the theoretical channels. However, these results are attractive but unrealistic. Therefore, the performance has been verified in Chapter 5 through realistic channel realizations extracted both from measurement data and from the experimental relay channel model. The method of how to get the realistic channels is described at the beginning of this chapter. Thereafter, for

25 1.2 CONTRIBUTIONS AND OVERVIEW 7 the measurement data based performance verification, both the impacts of the small scale channel propagation phenomena and the impacts of the large scale propagation phenomena have been investigated. For the relay channel model based performance verification, different positionings of a RS have been studied. Parts of the results in this chapter have been published in [10]. Finally, Chapter 5 summarizes this dissertation and gives an outlook to the future works.

26

27 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS 9 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS As discussed in Chapter 1, relay networks are distinguished from single-hop networks in the sense that they have intermediate RSs between a BS and a MS, by which the topology of relay networks becomes mesh-structured. With the mesh topology of relay networks, new technological problems have to be challenged in the channel modeling where the relationship between meshed links becomes to be one of the focus points. The current existing channel models [34] [35] [40] [36] [41] [42] concentrate on the modeling of single links. Even though for some of them it is stated that more sophisticated network structures like ad-hoc networks as well as relay networks are supported [41], the relationship between links is only partially modeled or unrealistically simplified. This chapter discusses comprehensively the requirements which are of crucial importance for the wideband MIMO relay channel modeling. According to these requirements, the most important channel metrics are highlighted, namely, the large scale correlation properties between links. The correlation includes both the intra-site correlation and the inter-site correlation. Furthermore, the methods, how to include the correlation properties of LSP in the whole channel modeling process, are discussed. The outline of this chapter is as follows. First of all, Section 2.1 summarizes the state-of-the-art research in the channel modeling and summarizes the main contributions of this chapter. Then, Section 2.2 describes the general channel modeling process. Thereafter, various MIMO channel modeling methods are summarized in Section 2.3. In Section 2.4, the existing state-of-the-art channel models are compared and the WIM, which provides the best match to the system requirements of relay networks, is chosen as the initial channel model. Thereafter, three kinds of correlation properties between links are proposed in Section 2.5 as the most important channel metrics of relay networks. Section 2.6 shows the methods of how to include and reproduce these correlation properties. 2.1 Background and own contributions Background and state-of-the-art Channel analysis for modeling started at the most beginning for the narrowband channel signal transmission where only the time domain channel variation has been considered, for example, the Rayleigh-fading channel model, the Rice-fading channel model, the Nakagami-fading channel model, and the Suzuki-fading channel model [43] [31] [13]. In the middle of 1990s, the wideband frequency resource is envisaged to be licensed to satisfy the

28 10 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS rapid demand on wireless communications. Thus, channel modeling needed to be expanded to reflect more accurately the wideband channel behaviors including the channel properties both in the time and frequency domains. At the end of 1990s, Foschini and Telatar published their results about the promising attractive MIMO capacity in [29] [30]. After the recognition of the importance of MIMO techniques, booming research activities have been shifted to MIMO systems. The intensive study on broadband MIMO channel models is one of the research areas where researchers focus on characterizing the MIMO propagation. The broadband MIMO channel models have been investigated in many communities: 3Generation Partnership Project (GPP) [34], Institute of Electrical and Electronics Engineers (IEEE) [35], European Cooperation in the field of Scientific and Technical Research (COST) [40] [36], and the most recent Wireless World Initiative New Radio (WINNER) project [44]. The classification of MIMO channel modeling approaches is discussed in [45] [46]. Generally speaking, they include the stochastic channel modeling [35] and the deterministic channel modeling [47] [48]. Both approaches described above mainly focus on reproducing the short-term MIMO wave propagation between two stations in a LSA 1, namely a link level channel modeling [5]. The large scale properties, such as the variation of propagation channels along a MS s moving trajectory from a LSA to another neighboring LSA and the correlation between MS s propagation channels to BSs/RSs, are not explicitly included. This indicatively means, the system level channel properties are not a central point of discussions. In [50] the authors modeled the channels of relay networks as a simple interconnection of intermediate links, again without considering the large scale properties. However, for relay networks, the large scale propagation properties are very important. It has been stressed in [51] [52] [39] [22] [53] that the performance of relay networks can be significantly improved by various advanced algorithms, such as cooperative relaying as well as virtual antenna arrays. These algorithms investigated in the literatures rely on both the availability of links and their large scale properties. To characterize the large scale behaviors of propagation channels, two correlation metrics; the intrasite auto- and cross-correlation and the inter-site correlation, are introduced. By introducing those correlation metrics, system level channel properties can be characterized. Furthermore, networks with a more sophisticated layout, such as relay networks, can be simulated. The intra-site auto- and cross-correlation models enable it to simulate single-cell multi-user applications as well as single-cell ad-hoc networks while the inter-site correlation enables multi-cell simulation like relay networks as well as Service Area (SA) [54]. Using the classical three-station structured relay networks [51] [52] [55] as an example, the intra-site auto- and cross-correlation can describe the similarity between the S-D 2 link and the S-R link while the inter-site correlation captures the correlation between the S-D link and the R-D link. In [56] [57], Fraile showed already the impacts of the intra-site auto-correlation properties of LSF on the system performances. It proves that the system-level performances in terms of throughput 1 The terminology LSA [49] is defined as the area where a channel can be regarded as wide-sense stationary. LSA is similar with the terminology segment in the WIM [1], drop in the 3GPP SCM [34]. 2 S-R-D stands for the relay channel in three-station structured relay networks, namely, S-R stands for the channel between the source and the RS, S-D the channel between the source and the destination, R-D the channel between the RS and the destination. The source is one from the MS and the BS, depending on the down-link or up-link transmission.

29 2.1 BACKGROUND AND OWN CONTRIBUTIONS 11 and Block Error Rate (BLER) are sensitive to the intra-site auto-correlation model of LSF. Furthermore, the importance of the inter-site correlation has been discussed in literature [58] [59] [60] [61] [62] [63] where it has been observed that the inter-site correlation plays a key role in system design like interference management, radio resource allocation, hand-over algorithm, power-control algorithm, cooperative process, and macro-diversity management. The recognition of the importance of these correlation properties pushes the research activities towards channel modeling. The experimental results of the intra-site correlation, based on measurement data, have been shown in [64] [65] [5] [66] [67] in various propagation scenarios. The intra-site cross-correlation were considered already in the WIM [1] [5] and in the 3GPP Spatial Channel Model (SCM) [34]. As an extension to the SCM [34], the WIM included the intra-site auto-correlation in the final deliverable [1] and in the channel simulation software [44]. Compared with the fruitful results of the intra-site analysis, the study to the inter-site correlation is countable. The representative literature about the inter-site correlation is [68] [69] [70] [71] [72]. All of them limit their study to the inter-site correlation of LSF in the macro-cell scenario except [73] due to the following facts: In first generation wireless communication systems SNR is the single key indicator when access the system performances. As shown in Eqn. 2.9, SNR is only decided by LSF with a defined PL model, a fixed transmit power, and a fixed transmit and receive antenna gain. This is the reason why LSF is the single LSP which was considered in the modeling of the inter-site correlation. Outdoor macro-cell systems with a high-elevated BS antenna were the most interesting application at the beginning of mobile communication since a large area can be covered with a single BS. This is the reason why only macro-cell scenarios were considered in the modeling of the inter-site correlation. The measurement equipment used in designing our legacy systems was limited in identifying the parameters related to the inter-site correlation modeling. Majority of the measurement campaigns performed for the purpose of the inter-site correlation modeling were for narrowband measurements. Thus, none of the channel properties in the delay domain can be captured. Furthermore, high resolution antenna arrays are not applied, either. Therefore, it is impossible to get sufficient channel information in the delay and spatial domains. In [73], the authors considered the inter-site correlation of Angular Spread (AS) and investigated the inter-site correlation properties in indoor scenarios. Regardless of these research activities in the inter-site correlation, a fixed inter-site correlation of LSF, which is typically 0.5, was considered in the simulation studies in [74] and in the 3GPP channel modeling as well [34]. In fact, this value is estimated from a few limited measurements and therefore it may not be useful in the general case. The WIM has investigated the inter-site correlation model theoretically. However, the inter-site correlation of LSP has been set to be 0 in the WIM, and excluded from the final deliverable [1] and from the channel simulation software [44].

30 12 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS Own contributions Since the correlation properties are crucial for the system level channel modeling in designing relay networks, the major contributions of this chapter were set at studying and investigating the correlation properties of LSP comprehensively. As a final result, a relay channel model will be proposed, based on the WIM. The main contributions of this chapter are listed as follows. A summary of MIMO channel modeling approaches. After presenting the general MIMO channel modeling process, MIMO channel modeling approaches including the stochastic and deterministic modeling methods are discussed. A comprehensive comparison among the state-of-the-art channel models. the state-of-the-art MIMO channel models, including the 3GPP SCM [34], the IEEE n channel model [35], and the COST 273 channel model [40] [36] as well as the WIM [44] [1], are compared in terms of modeling approach and model parameters as well as software implementation. According to the system requirements imposed on the relay channel modeling, the WIM is selected as an initial channel model. Proposal of the channel model for relay networks, based on the WIM. According to the system requirements and the network topology of relay networks, the goals of the channel modeling are discussed. Using the WIM as the initial channel model, the key channel metrics for relay networks have been proposed. They are the intra-site correlation and the inter-site correlation. Both of them reflect the long-term correlation properties between meshed links. These correlation properties have been studied for LSP, representing the large scale channel properties in the spatial, temporal, frequency, and power domains. Providing method of how to reproduce the channel correlation properties in the whole reproduction process. If the correlation properties of LSP are available, this chapter provides the methods of how to reproduce these channel metrics and of how to integrate these correlation properties into the whole channel modeling process. An example has been exploited to detail the whole reproduction process. 2.2 Channel modeling process Channel modeling is a cooperation work between propagation researchers and system designers [75]. The system designers provide the propagation researchers with the information on investigated systems and the requirements on channel models. At the end, the results of propagation modeling are output to the system designers. The general channel modeling process is shown in Fig The whole process is divided into 3

31 2.3 MIMO CHANNEL MODELING APPROACHES 13 System requirements System parameters Initial channel model Model parameters Modelling goals Channel parameters Processing of measurement data Measurement data Channel measurement campaign New model parameters Parameters regeneration Model parameters Model simulation Channel realizations System simulation & model validation Fig. 2.1: Channel modeling process phases. The first phase focuses on the initialization of a channel model and the modeling goals. The second phase aims at getting the channel parameters required both for the initial channel model and for the new modeling goals. In the third phase, the model parameters are reproduced and fed to the model to simulate channel realizations. The first phase starts with system requirements. After studying the investigated radio system, the required system parameters are returned. Based on those system requirements, the channel model, providing the best match to the considered system, is selected from existing channel models as the initial channel model. The model parameters are adopted from the initial channel model. However, the initial channel model can not fulfill all system requirements. Therefore, some new modeling goals are proposed, from which new model parameters are returned. To acquire the new model parameters as well as the initial model parameters, measurement campaigns are planned and then conducted at the second phase. Measurement data, stored in a massive memory disk array, is postprocessed, by which the model parameters can be extracted. Those parameters are input to the third phase where according to their statistical properties, the model parameters are reproduced by using a random number generating algorithm and a filtering algorithm. The reproduced model parameters are fed to the modeling algorithm for channel realization simulations. Using the channel realizations, system performance can be predicted. Furthermore, the channel model can be further verified by comparing the simulated channel realizations with those directly from measurement data. 2.3 MIMO channel modeling approaches For MIMO channel modeling, generally a distinction can be made between the stochastic and deterministic modeling approaches [45] [46]. Table 2.1 gives an overview of the categories. 1. Stochastic channel modeling stochastic modeling aims to describe the stochastic properties of channel parameters. It does

32 14 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS Tab. 2.1: Categories of MIMO channel modeling approaches MIMO channel modeling Stochastic channel modeling Deterministic channel modeling Correlation based [35] Geometric scatter based [34] Ray-tracing based [47] Measurement based parametric [48] Stored channels not depend on the site-specific description. In the stochastic section, two sub-categories are introduced, namely correlation based modeling and geometric scatter based modeling. (a) Correlation based channel modeling Under the assumption that the channel coefficients of a MIMO channel matrix H are complex Gaussian distributed, the first- and second-order moments of the channel coefficients can fully characterize the statistical behavior of a MIMO channel. Several models are developed based on the second-order statistics of a MIMO channel, such as the IEEE n channel model [35]. Correlation based channel modeling gives a limited insight into the propagation characteristics of a MIMO channel since no physical parameters, such as scatters, are used. The statistical correlation matrix R [76], expressing the correlation characteristics between any transmit and receive antenna pair, is exploited for the modeling purpose. R can be computed as R = E{vec(H) vec(h) H }, (2.1) where E{.} is the expectation operation, the superscript H denotes the complex conjugate transpose, H, a MIMO channel matrix, has a dimension of n Tx n Rx, and vec stands for the operation of stacking all columns of the channel matrix H into a vector. Using the correlation matrix R, a MIMO channel can be reproduced by vec(h) = R 1 2 G, (2.2) where G is a matrix with its elements being independent and identically distributed (i.i.d) zero mean complex Gaussian random variables. Thus, the correlation matrix R has a dimension: n Tx n Rx n Tx n Rx. As the number of antenna elements increases, the extreme large dimension of the correlation matrix makes the calculation more complicated and even impossible. A simplified Kronecker model [77] is proposed under the assumption that the transmit correlation matrix R Tx and the receive correlation matrix R Tx are independent, where R Rx = E{h i h H i },for i = 1,...,n Rx (2.3) and R Tx = E{((h jh )h j ) T },for j = 1,...,n Tx. (2.4) where h i stands for the i th column of H while h j stands for the j th row of H, and the superscript T denotes the complex transpose.

33 2.3 MIMO CHANNEL MODELING APPROACHES 15 The channel correlation matrix R can be approximated by the Kronecker product of R Tx and R Rx as R = R Tx R Rx, (2.5) and a MIMO channel can be reproduced by H = R 1 2 Tx GR T 2 Rx. (2.6) The principle of the Kronecker assumption could hold in the scenarios where scatters are uniformly distributed around both stations, for example a two-ring model [76]. However, this assumption often could not hold in realistic propagation. Mostly, the Kronecker model underestimates the correlation. The correlation based models are beneficial when analyzing the MIMO capacity [77]. (b) Geometric scatter based channel modeling Geometric scatter based channel modeling is based on the locations and properties of individual scatter in a given propagation environment. It is a physical modeling approach. The shape of a scatter area can be defined with respect to the type of scenario, for example, a one-ring model in the case when a BS is assumed to be highly elevated while a MS is surrounded by scatters [76], a two-ring model in the case when both a BS and a MS are surrounded by scatters [76], and an elliptical model [78]. The second variant of the geometric scatter based models describes the channel as a superposition of impinging waves [49]. Channel realizations are based on a tap delay line structure. Each path is a propagation path which is characterized by the spatial scatter distribution. 2. Deterministic channel modeling Deterministic channel modeling needs a precise description of the specific propagation scenario. Therefore, it is a site-specific channel modeling method. Three approaches are described in the following, namely ray-tracing based, measurement based parametric channel modeling, and stored channels. (a) Ray-tracing based channel modeling Ray-tracing based channel modeling is based on geometric optic approximations. An accurate description of the geometric environment and its electro-magnetic properties [47] [79] is required in ray-tracing. The Channel Impulse Response (CIR) at certain position is predicted by summing up the contribution from a large number of paths through the environment. Both the signal strength and the direction of each MPC can be computed. The more accurate the description of the environment, the better the match between a ray-tracing based simulation channel and a realistic channel. However, the computational effort is enormous. More general, there are many simplified ray-tracing based methods, by which the simulation complexity can be significantly reduced while providing a reliable channel simulation [80]. To predict the coverage in a certain environment and to compute the CIRs in an accurately defined environment, ray-tracing is the most popular method.

34 16 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS (b) Measurement based parametric channel modeling Measurement based parametric channel modeling [81] describes the propagation as a superposition of specular components [48]. First of all, the stored MIMO measurement data is processed by using a high-resolution channel parameter estimation algorithm like Estimation of Signal Parameters via Rotational Invariance Techniques (ESPRIT), Space Alternating Generalized Expectation maximization (SAGE), or Gradient based ML Parameter estimation algorithm (RIMAX) [82]. As a result, the parameters of each MPC can be estimated. The parameters, including the complex channel gain, delay, Angle of Arrival (AOA) and Angle of Departure (AOD) for azimuth and elevation, and Doppler frequency shift, represent the complex multipath CIR in the power, delay/frequency, spatial, and temporal domains. One of the aims of measurement based parametric channel modeling is to separate the influence of the antennas from the radio channel. A MIMO channel can be simulated by integrating the directivity of an antenna array into a multipath channel. Thus, an arbitrary antenna array can be included in synthesized channels. (c) Stored MIMO channels The stored MIMO channels from measurement campaigns are often used in the performance simulation for the verification purpose [10] [83] but rarely used to simulate radio channels directly. It is due to the fact that this approach is too site-specific, less flexible, and further measurement antenna array dependent. In stochastic channel modeling, a large number of random parameters can be a headache for a reliable channel simulation. Thus, several parameters will be fixed to reduce the system simulation complexity meanwhile to provide an acceptable accuracy. 2.4 Channel modeling process for relay networks The channel modeling for relay networks will be performed according to the process presented in Subsection 2.2. Each step will be described in detail in the following System requirements Nowadays, both wideband and MIMO are standard techniques. Thus, the channel of relay networks would be a wideband MIMO propagation channel. The wideband MIMO channel between any two stations in relay networks can be modeled as what has been conducted for single-hop networks, using the approaches described in Section 2.3. It is called a link level channel modeling where the short-term wave propagation of a single link is characterized and then modeled. However, a link level channel modeling is not sufficient for relay networks. It has been widely discussed and accepted that wireless communication systems benefit from relay concepts not only due to the availability of intermediate RSs, but also due to the meshed network topology. The ideas such as cooperative relaying as well as virtual antenna arrays [51] [52] [39] [22] [53] have been stimulated

35 2.4 CHANNEL MODELING PROCESS FOR RELAY NETWORKS 17 from the meshed network topology. Therefore, the channel modeling for relay networks should not only focus on modeling the channel between any two stations for short-term, but also the long-term properties between meshed two-station channels. Using the classical three-station structured relay networks as an example, the task of channel modeling for relay networks should not be limited to model the single S-D, S-R, and R-D channels. The correlation between them is of importance which should be treated as one of the focus points [5] in the channel modeling for relay networks Comparison among the state-of-the-art MIMO channel models To facilitate the channel modeling for relay networks, the state-of-the-art MIMO channel models, including the 3GPP SCM [34], the IEEE n channel model [35], and the COST 273 channel model [40] [36] as well as the WIM [44] [1], are compared. The one, providing the best match to the system requirements, will be selected as an initial channel model, based on which the channel model of relay networks will be proposed. Some comparisons have been done in Table A in Appendix [84] [41] [42] [44] [85]. It can be observed from Table A that all these channel models are based on the stochastic channel modeling approach. The IEEE channel model uses the Kronecker product of two correlation matrices as the modeling approach. The fact that it covers only indoor scenarios limits its application to relay networks. Therefore, it will not be considered as the basis for relay channel modeling. The modeling methodology of the remaining three channel models can be traced back to the COST 259 [40] framework: the ray-based spatial channel modeling. Cluster strategies are introduced which facilitate the channel modeling and reduce the modeling complexity. Both the WIM and the 3GPP SCM are a system level channel modeling and provide a software implementation. Furthermore, the former makes a significant improvement to the latter in terms of scenarios, bandwidth, and channel parameters. Therefore, the WIM is selected as the initial channel model of relay networks. The WIM has a top-down modeling philosophy where at first the system level propagation and then the link level propagation are reproduced. With this top-down structure as shown in Fig. 2.2, the WIM models firstly the multi-cell case by exploiting the inter-site correlation as a key channel metric. Thereafter, it shifts the focus to the single-cell case where the intra-site correlation is introduced. Thereafter, single-links are modeled by MPCs. Each path in MPCs comprises subpaths with a sum of sinusoids method. This top-down structure makes the WIM applicable for any kind of networks with various network topologies, for example relay networks. At the link level, the MPC parameters of a clustered delay line model (as shown in Fig. 2.3) describe the short-term propagation between two stations over a LSA. The parameters of each MPC are called small scale parameters. They include complex polarimetric path gain, departure (from the transmitter) and arrival (to the receiver) angles, and propagation delay. For each channel realization over a LSA, these MPC parameters are chosen randomly according to their Probability Density Function (PDF)s. The sub-paths of each MPC are generated by summing up equal-powered sinusoids [86], having the same delay, fixed AOA and AOD offsets, but different phases. Both in the WIM and in the 3GPP SCM, the small scale parameters are low-level model parameters. They are controlled by LSP which stay the same over a LSA and vary continuously from a LSA to another

36 18 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS System level to link level channel modelling Multiple-cell single-cell single-link Single-path Inter-site correlation Intra-site auto- and cross-correlation Multipath-component parameters Sum of sinusoids Fig. 2.2: The top-down modeling philosophy of the WIM neighboring LSA. The parameters, such as DS, AS, LSF, and K-factor, belong to the LSP group in the WIM while the LSP group in the 3GPP SCM includes only DS, AS, and LSF. Since the spatial, temporal, frequency, and power domains of a channel are most probably not independent, the intra-site cross-correlation between different LSP is introduced as a metric to quantify this interdependence. The intra-site cross-correlation has been considered already both in the WIM [1] [5] and in the 3GPP SCM [34]. MS1 AOA AOD BS1 Fig. 2.3: A link level channel modeling over a LSA Now shifting our observation from the single link short-term propagation illustrated in Fig. 2.3 to the multiple links long-term propagation illustrated in Fig. 2.4, two cases are considered. The one is that two MSs are connected to the same BS as shown in Fig. 2.4(a), the other is that two BSs communicate with the same MS as shown in Fig. 2.4(b). The question is raised: is there any long-term similarity among multiple radio links, between BS1-MS1 and BS1-MS2 or between BS1-MS1 and BS2-MS1, in the spatial, temporal, frequency, and power domains? Due to the existence of common scatters and reflectors among multiple links, the answer is definitively yes. Thereafter, we introduce the measures, the intra-site auto-correlation and the inter-site correlation to quantify those similarities. The intra-site auto-correlation is a channel metric which characterizes the similarity between multiple links from different MSs to the same BS as shown in Fig. 2.4(a) while the inter-site correlation is a channel metric which characterizes the similarity between multiple links from different BSs to the same MS as shown in Fig. 2.4(b). LSP are considered in calculating the two long-term correlations. By introducing the two correlation metrics: the intra-site auto- and cross-correlation and the intersite correlation, systems with a more complicated layout can be simulated. The intra-site autoand cross-correlation allows it to simulate channel realizations for single-cell multi-user applications

37 2.4 CHANNEL MODELING PROCESS FOR RELAY NETWORKS 19 and single-cell ad-hoc networks while the inter-site correlation allows a multi-cell simulation like relay networks and SA [54]. In 3GPP channel modeling, both the intra-site cross-correlation and the inter-site correlation have been considered. The intra-site cross-correlation of LSP has been investigated for LSF and DS as well as AS. But the inter-site correlation study is limited to LSF with a fixed correlation coefficient of 0.5. The intra-site auto- and cross-correlation has been proposed and implemented in the WIM. Furthermore, the WIM has investigated the inter-site correlation model theoretically. However, the inter-site correlation of LSP has been set to be 0 and excluded from the final software implementation in the WIM. Therefore, it can be concluded that the WIM supports both the link level and system level channel modeling. BS2 MS1 BS1 MS1 BS1 MS2 (a) The intra-site auto-correlation (b) The inter-site correlation Fig. 2.4: A system level channel modeling Modeling goals Although the correlation concepts have been proposed in the WIM, only the intra-site auto- and cross-correlation is included in the final deliverable [1] and implementated in the final channel simulation software [44]. Furthermore, some aspects of the correlation properties are not considered in the final deliverable due to insufficient measurement data. Moreover, the inter-site correlation is only introduced but not implemented in the WIM channel simulation tool. Since these correlations are crucial for the system level relay channel modeling, one of the major contributions of this chapter is to study and to investigate the correlation properties of LSP. As a final result, the channel model of relay networks will be proposed based on the WIM framework. The investigation includes the following aspects: 1. Except the parameters considered in the WIM, the following channel parameters will be included into the LSP group Cross-polarization Ratio (XPR); MIMO metric such as Normalized Parallel Channel Gain (NPCG) [77], which captures the spatial characteristics of a MIMO channel and gives an insight into the MIMO gain compared with a Single-Input Single-Output (SISO) channel. 2. AS in the elevation domain The WIM characterized the MIMO space only in the azimuth domain. In this chapter, the elevation domain will be also studied to provide a fully three dimensional description to the

38 20 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS MIMO space. 3. The intra-site auto- and cross-correlation The intra-site auto- and cross-correlation characterizes the correlation in two cases. The one is the case where two MS have a distance d MS from each other, to the same BS, the other is the case where one MS move a distance d MS, to the same BS [56]. Since both cases are equivalent in terms of the intra-site correlation modeling, only the first one is considered for the simplicity. The work conducted up to now focuses mostly on the relationship between the intra-site autoand cross-correlation and the distance between two MSs. A new dimension will be introduced to characterize the intra-site correlation, namely, the BS height. 4. The inter-site correlation In relay networks, the signals coming from different BSs/RSs are often correlated especially if common dominating scatters exist. For example, the inter-site correlation in the example scenario in Fig. 2.5(a) will be larger than the one in the example scenario in Fig. 2.5(b). BS2 BS1 BS2 BS1 MS1 MS1 (a) Example scenario (b) Example scenario Fig. 2.5: Examples of the inter-site correlation The first investigations of the inter-site correlation properties have been done only to one LSP, namely, LSF. The results in [71] [72] showed the dependence of the inter-site correlation on the angle seen from a MS to two sites: BSs/RSs. Other experimental studies showed that the inter-site correlation depends on the relative distance between two sites [63]. Besides these two factors, the distance difference between two links and the height difference between two sites will be considered as the impact factors in the inter-site correlation modeling. The inter-sector correlation, as a special case of the inter-site correlation, will also be investigated. 5. The BS/RS height The impact of the BS/RS height on PL models has been studied in [15]. But, no work has been done to investigate its impact on LSP. Thus, one of the focuses will study the impact of the BS/RS height on the statistical distributions of LSP and their correlation properties.

39 2.5 THE DEFINITION OF LSP AND THEIR CORRELATION PROPERTIES The definition of LSP and their correlation properties The following channel metrics are considered: Delay Window (DW) and DS as well as coherence bandwidth in the delay/frequency domain; Receiver (Rx), Transmitter (Tx), and NPCG in the spatial domain; PL, LSF, XPR, K-factor, and corner loss in the power domain. Those parameters are introduced into the LSP group except corner loss. 1. DW In a measured PDP, DW is defined as the time delay of multipath with power being larger than a cut level as shown in Fig The cut level is the threshold under which majority of the signal detected is the measurement thermal noise. 2. DS, coherence bandwidth, and AS In general, a spread is the root of the second central moment of a power spectral density. DS is computed over a delay power spectral density or a PDP while AS is calculated over an angular delay power spectral density or a Power Angle Spectrum (PAS). The coherence bandwidth is a measure of the frequency difference, at which the frequency correlation coefficient of a power frequency profile falls from 1 to a certain level. The coherence bandwidth, as a duality to DS, expresses the time dispersive properties of multipath channels in a LSA. The coherence bandwidth is inversely proportional to the DS τ rms [87] as: B c = 1 c τ τ rms, (2.7) where c τ ranges from 5 to 10 depending on the shape of a PDP [13] [88]. Furthermore, the coherence bandwidth has a lower bound [87]: B c arccos(c B )/2 πτ rms, (2.8) where c B stands for a selected coherence level. In Fig. 2.6, a PDP is shown as an example. In the same figure, DW, DS, mean delay, and noise level as well as a cut level are shown for the selected PDP. 3. PL, LSF, and SSF The received signal power P Rx in mobile communication is often modeled as a product of four factors (see Eqn. 2.9): the transmit power P Tx (in [dbm]), the distance dependent PL [13] (in [db]), log-normal distributed LSF (in [db]) [89], and SSF (in [db]). P Rx = P Tx PL + LSF + SSF. (2.9) PL characterizes the dependence of the signal attenuation on the distance between the transmitter and the receiver. The distance is expressed in meters. Note that the antenna gain might be included or excluded from a PL value. PL is often modeled as follows,

40 22 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS delay window RMS delay spread mean delay magnitude [db] db cut level noise level delay [µs] Fig. 2.6: An example PDP and the corresponding channel parameters PL = 10n PL log 10 d + B (2.10) where n PL is the PL exponent, d expresses the distance between the transmitter and the receiver in m, and B is the PL value when d is one meter. As a result of the multiplication of a large number of random attenuating factors along a propagation path, LSF has a log-normal distribution with a zero mean value and a σ 2 variance [89]. The variance σ 2 is expressed in db. SSF, characterizing the superposition of multipath propagation, is often modeled as a Rayleigh/Rice distribution. SSF is also called fast fading The PL model with antenna gains for the free space is given by PL = 10lg ( λ 2 ) WL G Tx G Rx (4π) 2 d 2 = 20lgd + 10lgG Tx + 10lgG Rx + B, (2.11) where λ WL stands for a wave length, G Tx and G Rx stand for the gains of the transmit and receive antennas, respectively. Without antenna gains the model is given by 4. XPR ( λ 2 ) PL = 10lg WL (4π) 2 d 2 = 20lgd + B. (2.12) XPR defines the ratio of the power between the co-polar propagation and the cross-polar propagation. 5. K-factor K-factor is defined as the ratio between the power of a direct LOS path and the power of the remaining MPCs in a LOS region. In Fig. 2.7, the first path stands for the power of a direct LOS path, while the sum of the rest paths means the power of the remaining MPCs. The ratio between them is computed as the K-factor value. 6. Corner loss [90]

41 2.5 THE DEFINITION OF LSP AND THEIR CORRELATION PROPERTIES 23 1 normalized linear PDP delay bin Fig. 2.7: PDP with a direct LOS path and the remaining MPCs Corner effect in actual urban environments is defined as the sudden decrease in the received power (in [db]) due to the change from the LOS propagation to the NLOS propagation when a MS turns around a corner. 7. MIMO metrics (a) Definition of a MIMO channel MIMO, as one of the significant breakthroughs in modern mobile communications, can improve the system capacity and the frequency efficiency [29] [30]. A MIMO channel, having n Rx receive antenna elements and n Tx transmit antenna elements, can be expressed by a matrix H with dimension n Tx n Rx. The system model of a MIMO transmission can be expressed as r 1. = h 1,1 h ntx,1. h ij. s 1. + n 1., (2.13) r nrx h 1,nRx... h ntx,n Rx s ntx n ntx where r i, 1 i n Rx, is the receive signal; h i,j, 1 i n Tx and 1 j n Rx, is the channel efficient; s i, 1 i n Tx, is the transmit signal; and n i, 1 i n Tx is the additive white Gaussian noise (AWGN) noise. With r = [r 1,...,r nrx ] T, s = [s 1,...,s ntx ] T und n = [n 1,...,n ntx ] T, Eqn can be reformed as r = H T s + n. (2.14) (b) Singular values After performing a singular value decomposition to the channel matrix H T, H T can be expressed as H T = U λ V H, (2.15) where λ is a n Rx n Tx matrix with diagonal elements being λ 1,...,λ j,...,λ min(ntx,n Rx ).

42 24 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS These elements are the singular values of the channel matrix H T. (c) The MIMO capacity Based on the definition of a MIMO channel, the MIMO capacity C MIMO can be calculated as C MIMO = log 2 (det (I nrx + ρ SNRH T H )). (2.16) where det stands for a matrix determinant operation, I nrx is an identity matrix with a n Tx dimension n Rx n Rx, ρ SNR is the SNR value of a MIMO channel. By replacing H T with Eqn. 2.15, Eqn can be rewritten as Eqn. C MIMO = log 2 (det (I nrx + ρ SNRU λ V H V λ H U H )) n Tx ( = log 2 (det U (I min(ntx,nrx) + ρ SNR λ λ H )) )U H min(n Tx,n Rx) = log 2 = i=1 min(n Tx,n Rx ) i=1 ( 1 + ρ ) SNR λ 2 i n Tx n Tx ( log ρ ) SNR λ 2 i. (2.17) n Tx 2.17 indicates that the capacity of a MIMO channel is the sum capacity of min(n Tx,n Rx ) parallel SISO channels. 1 + ρ SNR n Tx λ 2 i is the channel gain of the ith parallel SISO channel. The key idea of MIMO is to improve the date rate by spatial multiplexing in eigen mode where min(n Tx,n Rx ) different date streams can be transmitted over min(n Tx,n Rx ) SISO channels. In the best case, min(n Tx,n Rx ) times SISO capacity can be achieved. (d) NPCG As ρ SNR n Tx in Eqn becomes larger, 1 + ρ SNR n Tx λ 2 i ρ SNR n Tx λ 2 i. Consequently, C MIMO is increased linearly with min(n Tx,n Rx ). If min(n Tx,n Rx ) is fixed, the MIMO capacity is determined only by the singular value λ 2 i. To characterize the spread of MIMO singular values, the metric, NPCG, is introduced [36] [91]. NPCG is defined as follows, NPCG = min(ntx,n Rx ) i=1 λ 2 i max(λ 2 1,... (2.18),λ2 min(n Tx,n Rx )). Unlike the MIMO capacity, which is a function of the system SNR, NPCG is independent on the system SNR. However, NPCG can still characterize the gain achieved by MIMO. The NPCG value is lower-bounded by 1 and upper-bounded by min(n Tx,n Rx ). The upper-bound can be achieved if min(n Tx,n Rx ) equal powered parallel SISO channels exist: λ 2 1 =... = λ2 min(n Tx,n Rx ). 8. The intra-site correlation properties of LSP (a) The intra-site auto-correlation of LSP

43 2.5 THE DEFINITION OF LSP AND THEIR CORRELATION PROPERTIES 25 The intra-site auto-correlation of LSP indicates how fast the LSP values vary along the d MS distance of a MS, alternatively, how large the similarity of the LSP values is between two MSs separated with a distance being d MS as shown in Fig The MS1 d MS BS MS2 Fig. 2.8: The intra-site auto-correlation intra-site auto-correlation of LSP can be calculated as, ρ MS = C AB(d MS ) CAA C BB, (2.19) where C AB is the cross-covariance of A and B. A and B are the variables of the same LSP, but between two MSs being separated with a distance being d MS. As discussed in [64] [65] [5] [66], the spatial variation properties of LSP can be well modeled as an exponential decaying function expressing the change of LSP over the distance d MS, ρ MS = exp( d MS d decorr ). (2.20) This means, LSP of two links toward the same BS would experience correlations which are proportional to their relative distance d MS. This correlation is quantified by the de-correlation distance d decorr defined as the distance to which the correlation coefficient is dropped to e 1. (b) The intra-site cross-correlation of LSP The intra-site cross-correlation is defined as the correlation properties among LSP. It characterizes the correlation between various domains, such as the spatial, temporal, frequency, and power domains. The intra-site cross-correlation coefficients can be expressed with a correlation matrix [5] [67]. Each entity in the matrix is computed as, ρ AB = C AB CAA C BB, (2.21) where C AB is the cross-covariance of two LSP, A and B, of the same MS. Since ρ AB = ρ BA, the cross-correlation matrix is symmetric. 9. The inter-sector correlation properties of LSP 3 3 If a channel model is antenna-independent, the inter-sector correlation is equal 1. If each sector of a BS is treated as a single BS, the inter-sector correlation merges into a special case of the inter-site correlation, where two BSs are the two sectors of the same BS.

44 26 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS The inter-sector correlation is defined as the correlation of the LSP of a MS which can get signal from two different sectors of the same BS. Let s define the 3-dB main beam widths of sector 1 and sector 2 to be θ s1 and θ s2 (as shown in Fig. 2.9), respectively. The inter-sector correlation of LSP is divided into two classes. The one is the correlation in the overlapped main beam area θ OL where a MS is in the main beams of two sectors, namely, θ OL = θ s1 θ s2 ; (2.22) the other is the correlation in the θ s1 θ s2 area or in the θ s1 θ s2 area where a MS is in the main beam of one sector but in the side beam of another sector. The inter-sector correlation can be expressed as, ρ AB = C AB CAA C BB, (2.23) where C AB is the cross-covariance of A and B. A and B are the variables of the same LSP, but from two sectors of the same BS. Fig. 2.9: The inter-sector correlation 10. The inter-site correlation properties of LSP 4 First of all, an simple example will be provided as follows to show the impact of the inter-site correlation of LSF on the system hand-over management. Handover rule: a MS communicates with the BS which provides a larger received power. Ignoring SSF, the received power can be calculated according to Eqn With the parameters defined in Table 2.2, the difference in the received power at a MS from two BSs is shown in Fig where only the PL model is considered. The larger the k d value, the larger the difference in the received power between two links. Under the hand-over assumption, the 4 Only the correlation of the same LSP from two links is considered.

45 2.5 THE DEFINITION OF LSP AND THEIR CORRELATION PROPERTIES 27 BS1 d 1 MS1 d BS d 2 BS2 Fig. 2.10: A hand over scenario Tab. 2.2: Hand-over parameter setup n PL 3.5 B 38.4 σ 8 db d 1 and d 2 d 2 = k d d 1 d 1,d 2 in [km] Path loss difference [db] n=3.5 n= k d Fig. 2.11: Received power difference (in [db]) from two BSs with different PL exponents

46 28 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS probability of hand-over is shown in Fig with the inter-site LSF correlation coefficient as a parameter. The inter-site LSF correlation coefficient varies from 0 to 0.8 with a step of 0.2. When a MS has the same distance to two BSs, the probability of handover at the MS is 50 %, 0.5 Handover Probability ρ = 0.8, 0.6, 0.4, 0.2, k d Fig. 2.12: Handover probability Based on the studies conducted to the inter-site correlation of LSF, it is proposed that the inter-site correlation of LSP can be expressed as, ρ BS = f(d diff,θ,d BS,h diff ), (2.24) min(d where d diff = 10log 1,d 2 ) 10 max(d 1,d 2 ) (ref. Fig. 2.10) stands for the distance difference between two links, θ is the angle of arrival difference from two BSs/RSs to the same MS, namely the angle seen from a MS to two BSs/RSs, d BS means the distance between two BSs/RSs, and h diff is the height difference between two BSs/RSs. The proposed new model includes and summarizes the ideas in [92] [71] [72] [63]. Additionally, two new parameters, d BS and h diff, have been introduced. 2.6 Usage of the correlation models of LSP Under the assumption that the experimental models of the correlation properties of LSP are available, it is important to accurately regenerate the correlation properties in the relay channel modeling process. The method will be addressed in this section. Recall the top-down modeling philosophy of the WIM presented in Fig. 2.2 in Subsection 2.4.2, the channel modeling for relay networks should focus on the first two parts: multi-cell and single-cell. The rest parts stay the same. The regeneration process is shown in Fig First of all, an algorithm, called weighted sums of independent Gaussian random processes, is used to represent the inter-site correlation of LSP. Then, a two-dimensional filter based method [93] will be used to introduce the intra-site auto-correlation.

47 2.6 USAGE OF THE CORRELATION MODELS OF LSP 29 Weighted sums of independent Gaussian random processes Inter-site correlation Two-dimensional filter based method Intra-site autocorrelation Linear transformation Intra-site crosscorrelation Fig. 2.13: Regeneration process and algorithms based on the correlation models of LSP Finally, a linear transformation is conducted to reproduce the intra-site cross-correlation. By doing this, the WIM can be adapted for a full relay network support. Furthermore, the WIM and the relay channel model are compatible. This is the case since LAYOUTPAR is one of the input parameters to start channel simulations in the WIM. This parameter provides two-dimensional coordinate informations of BSs/RSs and MSs [1]. The regeneration process is described in detail as follows. Assume that there are n BS one-sector BSs, n MS MSs, each of which has connection to n l (i) BSs (1 i n MS ), and n LSP LSP. The geometric area where all BSs and MSs locate has the dimension m x m y with a resolution of 1 [m]. 1. Generate random Gaussian processes n LSP ( n MS i=1 n l(i)) random Gaussian matrices with a zero mean value and a unit variance are generated. Each matrix has a dimension m x m y. Each MS-BS link has n LSP matrices. 2. The inter-site correlation Since the geometric structure of the simulation system is known, the inter-site correlation coefficient ρ BS of each LSP can be calculated according to Eqn To each MS, the inter-site correlation of each LSP can be expressed as a n l (i) n l (i) symmetric matrix. Then, performing Cholesky decomposition to the inter-site correlation matrix, a lower triangular matrix will be returned. By weighting n l (i) m x m y independent Gaussian random matrix with this lower triangular matrix, the inter-site correlation is introduced to one LSP. Using the same method, the inter-site correlation of the rest LSP and the rest MSs can be introduced. 3. The intra-site auto-correlation In literature, several design methodologies have been reported to simulate the intra-site autocorrelation properties. The majority of these methods can be classified into one of the following three categories [94] [95]. (a) The Sum-Of-Sinusoids method

48 30 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS This method is based on the fact that a Gaussian random process can be expressed as an infinite Sum-Of-Sinusoids with random phases and properly selected frequencies. In practice, a finite number of sinusoids can be used to approximate the Gaussian process so that the computational complexity can be reduced while keeping the final results comparable. This method is frequently used to simulate a correlated fast-fading process [86] [43] [31]. As mentioned in Subsection 2.4.2, Sum-Of-Sinusoids method has been used to model each MPC. (b) The filter-based method With this method, channel parameters are generated by means of filtering a zero-mean white Gaussian noise process. The frequency response of the filter is the square root of the Power Spectral Density (PSD) of a Gaussian process [94]. (c) Markov processes based method This method is often used to model radio link quality, such as the SNR and the BER, instead of being used to characterize a physical propagation. Due to the fact that the two-dimensional coordinate information is provided in the WIM, it is convenient to use a two-dimensional filter based method to reproduce the intra-site autocorrelation in the channel simulation. Recall the intra-site auto-correlation model proposed in Eqn in Section 2.5, the parameter d MS is defined as the distance between two MSs. Due to the fact that d MS = d 2 x + d 2 y = (x 1 x 2 ) 2 + (y 1 y 2 ) 2 (2.25) where (x 1,y 1 ) and (x 2,y 2 ) are the positions of two MSs in a two-dimensional Cartesian coordinate system, Eqn can be rewritten as ( ρ MS = exp( d ) MS (x1 x 2 ) ) = exp 2 + (y 1 y 2 ) 2. (2.26) d decorr d decorr Now the intra-site auto-correlation model is extended to a two-dimensional coordinate system. To extract the two-dimensional filter function from the intra-site auto-correlation model, following steps are conducted based on the theory that the Fourier transform of an autocorrelation function is the same as the square of the Fourier transform of the related CIR, (a) generate a two-dimensional m x m y matrix ρ MS to describe the intra-site auto-correlation in a coordinate system, the point with a correlation coefficient being 1 is in the middle of the map. (b) Perform a two-dimensional discrete Fourier transformation to ρ MS : fft(ρ MS ). (c) Get the root square: fft(ρ MS ) fft(ρms ) (d) Perform a two-dimensional inverse discrete Fourier transformation ifft( ). Note that a proper normalization should be performed to keep the variance to be one. then a two-dimensional filtering function f(x, y) is returned. The output at the second

49 2.6 USAGE OF THE CORRELATION MODELS OF LSP 31 step is filtered with this two-dimensional filter to introduce the exponential intra-site autocorrelation. 4. The intra-site cross-correlation After filtering, the LSP values at the MSs positions are selected for the current channel simulation while the values over the whole map can be saved as a look-up table for the coming channel simulation with the same setup. Now each MS-BS link has n LSP LSP values ζ. The intra-site cross-correlation is generated for each MS-BS link by a linear transformation, ˆζ = ρ AB ζ, (2.27) where ρ AB is the symmetric intra-site cross-correlation matrix described in Eqn in Section 2.5. Now we use an example scenario to illustrate the process described above. In the example scenario, n BS = 2, n MS = 2, n LSP = 2, m x = m y = 400, MS1 communicates both with BS1 and with BS2 while MS2 communicates only with BS2. The geometric distribution of MSs and BSs in a two-dimensional map is shown in Fig BS1 MS1 BS2 400m 400m MS2 Fig. 2.14: Geometric distribution of MSs and BSs 1. Generate matrices with a random Gaussian distribution, of which the mean value is zero and the variance is one. Each LSP matrix of the MS1-BS1 link and the MS1-BS2 link g i (1 i 2) has the dimension where 2 stands for two BSs. 2. From the geometric map, the inter-site correlation coefficients between the MS1-BS1 link and the MS1-BS2 link are for example 0.3 and 0.8 for LSP1 and LSP2, respectively. Correspond ingly, the inter-site correlation matrices can be expressed as and for

50 32 2. CHANNEL MODELING METHODOLOGY OF RELAY NETWORKS LSP1 and LSP2, respectively. After performing Cholesky decomposition to these two matri- 1 0 ces, and are returned. Weighting g 1 with the first matrix and g 2 with the second matrix, the inter-site correlation is introduced to LSP1 and LSP2. 3. Let s assume that LSP1 and LSP2 have, respectively, a de-correlation distance of 10 [m] and 50 [m]. Two two-dimensional matrices are generated to represent the intra-site autocorrelation. Using the steps addressed above, the filtering function f(x, y) can be extracted from the exponential intra-site auto-correlation functions. Convolving the filter with the LSP matrices, we can get the LSP values for each MS-BS link. 4. Selecting the LSP values at each MS position and saving the values over the whole map, we can get a vector with n LSP values for each BS-MS link. Vectors ζ 11, ζ 21, and ζ 22 are returned. Let s assume the intra-site cross-correlation between LSP1 and LSP2 is 0.4. Then the final LSP values at the MS position for each MS-BS link ζ ˆ 11, ζ21 ˆ, and ζ ˆ 22 can be calculated as ˆζ = ζ. (2.28) These LSP are fixed at each LSA. Their change from one LSA to a neighboring LSA can be observed from the intra-site auto-correlation. Figure 2.15 shows the change of one LSP over the whole area with the inter-site correlation ρ BS and the de-correlation distance d decorr as parameters. Three cases are compared in Fig. 2.15: 2.15(a) ρ BS = 0.8 and d decorr = 50 [m]; 2.15(b) ρ BS = 0 and d decorr = 50 [m]; 2.15(c) ρ BS = 0 and d decorr = 100 [m]. The similarity between two subplots in 2.15(a) is obvious due to the 0.8 inter-site correlation. Whereas, there is almost no similarity between two subplots in 2.15(b) and 2.15(c) where ρ BS = 0. Comparing 2.15(b) and 2.15(c), it is observed that the larger the intra-site auto-correlation, the slowlier the LSP values change over distance.

51 2.6 USAGE OF THE CORRELATION MODELS OF LSP (a) ρbs = 0.8, ddecorr = 50m (b) ρbs = 0, ddecorr = 50m (c) ρbs = 0, ddecorr = 100m Fig. 2.15: Distribution of the LSP values over the whole area with the inter-site correlation ρbs and the de-correlation distance ddecorr as parameters

52

53 3. EXPERIMENTAL CHANNEL EVALUATION AND MODELING EXPERIMENTAL CHANNEL EVALUATION AND MODELING Recall the channel modeling process shown in Fig. 2.1 in Section 2.2, the first and the third phases have been discussed in Chapter 2. This chapter focuses on the second phase: channel measurement campaigns and measurement data post-processing. This phase aims at extracting the channel parameters from measurement data for the relay channel modeling. Both the extraction algorithms and the processing procedure are presented. However, with the same algorithms and procedure, the experimental results may be different. The reason is that the setting parameters during postprocessing could be different. To obtain reasonable experimental results, the impacts of the setting parameters on channel parameters are investigated. Appropriate values of the setting parameters are selected for the final measurement data analysis. The LSP values and their correlation properties are highlighted for both the urban and indoor scenarios. This chapter is structured as follows. First of all, Section 3.1 gives firstly an outlook to the stateof-the-art research activities in the channel modeling for relay networks and summarizes the main contributions of this chapter. Thereafter, Section 3.2 introduces the scenarios which suit for relay applications. The relay measurement campaigns performed in two scenarios: indoor and outdoor urban, are described in Section 3.3. The same section provides a short description to the working principle of the channel sounder used. The extraction procedure and algorithms of the channel parameters are presented in Section 3.4 while Section 3.5 shows the experimental results of the LSP values and their correlation properties. 3.1 Background and own contributions Background and state-of-the-art Gudmundson proposed in [92] to model the intra-site auto-correlation function of LSF as an exponential decaying function, where the correlation coefficient is a function of the distance between two MSs. This model has been confirmed many times by the experimental results based on measurements [72] [64] [65] [3] [6]. As an extension, the authors in [64] [3] have investigated the intra-site correlation of LSF and DS as well as AS. The experimental decaying model is extensively used in mobile communications testbeds and simulation studies [96] [74]. Compared with the fruitful results for the intra-site correlation modeling, there is still no commonly acceptable model for the inter-site correlation. In 1978, Graziano published his experimental results of the inter-site correlation of LSF based on 900MHz channel measurements [68]. In his paper, he showed that the inter-site LSF correlation tends to be a function of the angle of arrival difference (θ

54 36 3. EXPERIMENTAL CHANNEL EVALUATION AND MODELING in Fig. 2.10); the smaller the angle, the larger the correlation. After his work, Mawira proposed in 1992 that the inter-site LSF correlation can be modeled as a linear function of the angle of arrival difference [69] which is in line with the experimental observation in [68]. In 1999, Sorensen confirmed Mawira s and Graziano s observation in [70] that the inter-site LSF correlation depends on the angle of arrival difference θ. In his paper he proposed to model the correlation as a piecewise function of θ. However, Perahia et. al observed that no association between the inter-site LSF correlation and θ has been found and the correlation is relatively low both in rural and in suburban environments [71]. Thereafter, Weitzen et. al doubted the conclusion drawn by Graziano with the argument that no enough data sets were available in the measurements used in [68]. Based on Weitzen s relatively massive data sets, he observed in his paper [72] that the average correlation coefficient is smaller than 0.2 even when the angle of arrival difference θ is small. The reason is that the distance difference between two links (d 1 and d 2 ) is very large. In the case of a large distance difference while small θ, the common objects causing shadowing between two links are still very rare. As a consequence, a MS experiences no similar LSF from two BSs. This means that both the distance difference between two links and the angle of arrival difference are decisive factors in the modeling of the inter-site correlation. The first investigation of the inter-site correlation properties has been done only to one LSP, namely LSF. The results in [71] [72] showed the dependence of the inter-site correlation on the angle seen from a MS to two sites: BSs/RSs. Experimental studies in [63] showed that the inter-site correlation depends on the relative distance between two sites. The results are estimated from a few limited measurements and therefore it may not be useful in a general case. The first results of the inter-site correlation in indoor scenarios have been presented by Jalden and Hong in [6]. The results showed that in the indoor case, the inter-site correlation is rather high for one measurement route but quite low for another route [6]. The WIM has investigated the inter-site correlation model theoretically. However, the inter-site correlation of LSP has been set to be 0 and excluded from the final software implementation in the WIM Own contributions Through the analysis on measurement data gathered from the urban and indoor scenarios, the experimental results of the statistical distributions of LSP and their correlation properties are presented in this chapter. The main contributions of this chapter is listed as follows: Except LSF, DS, and K-factor as well as AS, the channel parameters such as XPR and the MIMO metric NPCG, have been considered as LSP. Investigating the dependence of the LSP and their intra-site correlation properties on the BS height [7]. The intra-site correlation properties have been normally modeled as a function of the distance between two MSs. During the urban measurement campaign, the BS antenna has been set to two heights: 10 [m] and 16 [m], representing an urban micro- and macro-cell scenario,

55 3.2 PROPAGATION SCENARIOS 37 Tab. 3.1: Propagation scenarios for relay network applications Scenario Definition v MS h BS h RS h MS [km/h] [m] [m] [m] U1 Micro-cell urban 1 50 below rooftop U2 Macro-cell urban 1 50 above rooftop I1 Indoor corridor respectively. Therefore, the impact of the BS/RS height on the statistical distributions of LSP and their intra-site correlation properties can be studied. Experimental results on the correlation behaviors between two sectors of a BS. Both in the urban measurement campaign and in the indoor measurement campaign, the BS antenna array has been set to two orientations, angle of which is 90 [deg.] and 180 [deg.], respectively. A high inter-sector correlation can be observed in the overlapped main beam area. Experimental model of the inter-site correlation. The dependence of the inter-site correlation properties of LSP on the following parameters has been studied based on measurement data, distance difference between two links, angle of arrival difference from two BSs/RSs to a MS, the distance between two BSs/RSs, and the height difference between two BSs/RSs. 3.2 Propagation scenarios The propagation scenarios defined for relay network applications are summarized in Table 3.1. A description of each scenario is provided in the following subsections Urban scenarios U1 scenario: urban micro-cell U1 scenario defines the environments where both a BS antenna array and a RS antenna array are deployed lower than surrounding buildings. The BS antenna array is normally placed at the main street while the RS antenna array is located at the crossroad to provide coverage to a shadowed street, being perpendicular to the main street. This scenario covers both LOS and NLOS propagation conditions. The environment is defined as a Manhattan-like grid scenario, one of the most promising relay application scenarios [17] U2 scenario: urban macro-cell U2 scenario is distinguished from U1 scenario because the BS antenna array is above surrounding buildings. In this scenario, the BS antenna array provides a relatively large coverage due to its

56 38 3. EXPERIMENTAL CHANNEL EVALUATION AND MODELING height. The RS antenna array is expected to extend coverage or to provide coverage to a shadowed area. Both LOS and NLOS propagation conditions are considered Indoor scenarios I1 scenario: indoor corridor I1 scenario represents a typical office environment. The BS antenna array is assumed to be in corridor. Thus, it is a LOS propagation in the corridor where the BS is located while a NLOS propagation in the perpendicular corridors. Both the BS and RS antenna arrays have the same height. 3.3 Relay measurement campaigns Relay measurement campaigns aim to capture the correlation properties of LSP. The scenarios defined in Section 3.2, U1, U2, and I1, are considered in measurement campaigns. Before going detailed to each measurement campaign in Subsection and Subsection 3.3.3, Subsection gives a short introduction to the working principle of the channel sounder equipment used [97] Broadband radio channel sounding: technique and equipment The channel sounder used [98] [97] measures MIMO CIRs between n Tx transmit antenna elements and n Rx receive antenna elements. n Tx n Rx MIMO sub-channels are measured sequentially by the synchronous switching of the transmit and receive antenna elements. Figure 3.1 shows the switching time frame of the sequential MIMO channel sounding with a 3 4-antenna configuration. A SISO measurement has been conducted at each measurement time by sending a periodic multi-sine excitation signal [98]. This approach is well known from the frequency domain system identification in measurement engineering. In communication engineering this signal may be called Multi-Carrier Spread Spectrum Signal (MCSSS). Given the time length of a SISO CIR to be τ max, the total time length of a MIMO snapshot is 24τ max (i.e. 2n Tx n Rx τ max ). The factor 2 comes from the fact that a guard time is inserted after each SISO measurement to avoid switching transients. During measurements, the SISO CIR length τ max is adjustable, depending on measurement environments and on the distance between the transmitter and the receiver. The larger the distance, the larger the value τ max should be. Normally, τ max is larger in outdoor scenarios than in indoor scenarios. Two MIMO channel sounders, ATM sounder and HyEff sounder, are manufactured by the MEDAV GmbH, Germany [99]. The key technical parameters of these two sounders are summarized in Table. 3.2 [98]. The high resolution antenna arrays used in measurements are optimized to estimate the spatial dimension of a MIMO CIR with high resolution estimation algorithms, e.g. RIMAX [82] [98] [100]. Figure 3.2 summarizes the measurement antenna arrays for the high resolution parameter estimation, of which a Polarimetric Uniform Linear Array (PULA) is used as the BS/RS antenna array while either a Polarimetric Uniform Circular Patch Array (PUCPA) or a Stacked Polarimetric

57 3.3 RELAY MEASUREMENT CAMPAIGNS 39 Tx Rx Tx Rx Tx switcher Rx switcher SISO-CIR length Time Total Snapshot Time Length of one MIMO-CIR Fig. 3.1: MIMO channel sounder switching scheme Tab. 3.2: The key technical parameters of the MIMO channel sounder equipments Name Carrier frequency Bandwidth τ max Tx power RF sensitivity GHz [MHz] [µs] [W] [dbm] HyEff ATM 4.5/ /2/10-92 Uniform Circular Patch Array (SPUCPA) is used as the MS antenna array. A detailed 3-Dimension (3D) description to the antenna pattern of these antenna arrays is given in [98] Micro-cell to macro-cell urban scenario: U1 and U2 The urban measurement campaign has been performed, using the HyEff channel sounder, in Aug in the center of Ilmenau, Germany. The measurement area can be characterized as a canonical small urban scenario with 1-3 storeys stone buildings at the both sides of streets. Figure 3.3 shows the environment seen from BS3 and BS6, 16 [m]elevated from the ground [7] [8]. The average rooftop level is estimated to be between [m]. Street widths do not exceed the root-top heights. The positions of the RS and BS antenna arrays, PULA, have been shown in Fig The dashed curves in Fig. 3.3 stand for the moving routes of the MS, SPUCPA. Note that only part of the measurement routes are shown in Fig All measurement routes and the BS/RS positions have been shown in Fig. 3.4 in a 2-dimensional Ilmenau city map. The black circles in Fig. 3.4 stand for the positions of the BS/RS while the dashed curves are the moving routes of the MS. The arrows in Fig. 3.4 show the orientation of the BS/RS. The BS antenna array has two orientations: BS2/BS3 and BS5/BS6, the angle of which is 90 [deg.], θ OL = 90 [deg.]. Furthermore, the BS antenna array has been set to two different heights, 10 [m] (BS2/BS5) and 16 [m] (BS3/BS6) with the help of a lifting ramp while the height of the RS antenna array is fixed to 3 [m]. Increasing

58 40 3. EXPERIMENTAL CHANNEL EVALUATION AND MODELING Name Directivity No. of elements Inter-element space Polar. Picture BS/RS PULA 120 o WL V H MS PUCPA 360 o WL V H SPUCPA 360 o 4 x WL V H Fig. 3.2: Antenna arrays used for a high resolution parameter estimation RS1 BS2, BS5 RS2 RS3 BS3, BS6 MS Fig. 3.3: Ilmenau downtown seen from the 16 [m] BS antenna array the BS antenna height from 10 [m] to 16 [m], both the U1 and U2 scenarios are covered in the campaign. The detailed setup information of the channel sounder and of the transmit and receive antenna arrays is given in Table Indoor corridor scenario: I1 The indoor measurement campaign has been performed in Feb inside the Helmholtz building, at the university campus of TU-Ilmenau, Germany. Measurements were conducted with 8 different BS positions. At two BS positions, two different antenna broadside orientations are used, BS2,BS3 and BS7,BS8, the angle of which is 180 [deg.] as shown in Fig. 3.6, θ OL = 0 [deg.]. The BSs were placed 2.4 m above floor ground. Most of the BSs are located in the corridor area except BS5/BS6. Measurements routes have been chosen along corridors and in two seminar rooms (shown as dashed curves in Fig. 3.6). The black circles in Fig. 3.6 stand for the positions of the BSs. The arrows show the orientation of the BSs.

59 3.3 RELAY MEASUREMENT CAMPAIGNS 41 RS2 BS5/BS6 BS2/BS3 RS1 RS3 100m Fig. 3.4: Measurement routes in a 2-dimensional Ilmenau city map

60 42 3. EXPERIMENTAL CHANNEL EVALUATION AND MODELING Tab. 3.3: TU-Ilmenau relay measurement campaigns: U1, U2, and I1 Measurement campaign Institution TU-Ilmenau Scenario U1, U2 I1 Location Ilmenau downtown Helmholtzbau, TU-Ilmenau campus Date/Time Aug Feb Measurement setup MIMO Channel sounder Type RUSK HyEff [97] RUSK HyEff [97] center frequency [GHz] Bandwidth [MHz] CIR length [µs] No. of sub-channels Transmit antenna array (BS/RS) Name (type) PULA PULA Height [m] 3,10, Azimuth [deg.] Tilt [deg.] -5 at the BS only 0 Mobility [km/h] 0 0 Receive antenna array (MS) Name(type) SPUCPA PUCPA Height [m] Tilt [deg.] 0 0 Mobility [km/h] 5 3 BS9 BS7 BS1 MS Fig. 3.5: Photos of the Helmholtzbau measurement environment